A study finds that as few as 250 malicious documents can produce a “backdoor” vulnerability in an LLM, regardless of model size or training data volume
regardless of the size of the model or its training data. Data-poisoning attacks might be more practical than previously believed. [image] @anthropicai : Previous research suggested that attackers mig...
Sources: OpenAI overhauled its security, adding biometric checks in its offices and isolating sensitive info, to protect IP such as model weights from espionage
Artificial intelligence group has added fingerprint scans and hired military experts to protect important data
Geoffrey Hinton fears AI companies are under-investing in safety research, embracing military usage, and sharing model weights while pushing for less regulation
Nobel laureate Geoffrey Hinton, often called a “godfather of artificial intelligence,” spoke with Brook Silva-Braga …
OpenAI calls DeepSeek “state-controlled” and recommends that the US ban “PRC-produced equipment and models that violate user privacy and create security risks”
https://techcrunch.com/... Threads: Vishvanand Subramanian / @vishvanands : trying hard to steelman this position from openai but unless it's possible to hide malware in the model weights, what exactl...
When asked in an AMA if OpenAI would release model weights and research, Sam Altman said “we are discussing” and “it's also not our current highest priority”
www.theguardian.com/commentisfre... [image] Mastodon: Bryan Lawrence / @bnlawrence@mastodon.nz : “For us little people, the choice seems to be between being data-jacked and screwed over by the undemo...
When asked in an AMA if OpenAI would release model weights and research, Sam Altman said “we are discussing” and “it's also not our current highest priority”
CEO of the ChatGPT maker says his company has been ‘on the wrong side of history’ with open-source software
An evaluation of six frontier AI models for in-context scheming when strongly nudged to pursue a goal: only OpenAI's o1 was capable of scheming in all the tests
It presents a new safety challenge that OpenAI is trying to address. — techcrunch.com/2024/12/05/o... Anders Sandberg / @arenamontanus : In an IVA discussion on AI yesterday evening professor Kristi...
OpenAI, Microsoft, Meta, Google, Amazon, Anthropic, and Inflection make voluntary AI promises to the White House, like cybersecurity investment and watermarking
White House Pranav Dixit / Business Today : OpenAI, Google, Meta, Amazon and others pledge to watermark AI content for safety Ryan Morrison / Tech Monitor : White House secures AI safety commitment As...